The evaluation is fast because it's all prefill computation with only a single token of inference. Ditto cost, you're paying 100% input costs and nearly zero output. There really isn't any architectural magic to Jev, it's just a straightforward application of normal LLM tech with some good marketing.

I mean, Jev is also probably cheaper because it's a rather small model (or at least, I suspect it is based on the overall level of intelligence it demonstrates) so that helps make it cheap too.